Learnable Spatial-Temporal Positional Encoding for Link Prediction
Authors: Katherine Tieu, Dongqi Fu, Zihao Li, Ross Maciejewski, Jingrui He
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 5, we evaluate the empirical performance of L-STEP extensively, including (1) the comprehensive effectiveness comparison with 13 classic datasets, 10 algorithms, 2 learning settings (transductive and inductive), and 3 sampling strategies, where our L-STEP performs the best across the board, (2) the verification of the role of MLPs and transformers upon L-STEP, (3) the learning robustness of L-STEP in terms of different initial positional encoding inputs, (4) parameter analysis, running time comparison, and ablation studies, and (5) the leading performance in the large-scale TGB open benchmark (Gastinger et al., 2024). |
| Researcher Affiliation | Collaboration | 1University of Illinois Urbana-Champaign 2Meta AI 3Arizona State University. Correspondence to: Jingrui He <EMAIL>. |
| Pseudocode | Yes | Appendix B: Pseudo-code of L-STEP |
| Open Source Code | Yes | Our code is available at https: //github.com/kthrn22/L-STEP. |
| Open Datasets | Yes | We assess the ability of L-STEP in performing link prediction with 13 datasets covering various domains and collected by (Poursafaei et al., 2022): Wikipedia, Reddit, MOOC, Last FM, Enron, Social Evo., UCI, Flights, Can. Parl., US Legis., UN Trade, UN Vote, and Contact. Details about the dataset statistics are shown in Appendix G. We compare L-STEP with 8 state-of-the-art baselines... The large-scale datasets with pre-defined splits are publicly available at TGB Benchmark (Gastinger et al., 2024) |
| Dataset Splits | Yes | Across all 13 datasets, training/validation/testing sets are following the standard library (Yu et al., 2023) by chronological splits with ratios 70%/15%/15%. |
| Hardware Specification | Yes | The experiments are coded by Python and are performed on a Linux machine with a single NVIDIA Tesla V100 32GB GPU. |
| Software Dependencies | No | The experiments are coded by Python and are performed on a Linux machine with a single NVIDIA Tesla V100 32GB GPU. (No specific Python version or library versions mentioned.) |
| Experiment Setup | Yes | We first report the configuration and hyper-parameters that are unchanged for all 13 datasets: Dimension of time encoding: d T = 100. Dimension of node encoding: d N = 172. Dimension of edge encoding: d E = 172. Dimension of positional encoding: d P = 172 (only for Social Evo., d P = 72). Hyper-parameters for time encoding function: α = 10, β = 10. Weight of negative samples in positional encoding loss: αneg = 0.3. Weight of positional encoding loss in objective loss function of L-STEP αpe = 0.5. ... We leverage the Adam optimizer with learning rate of 0.0001. For a more detailed description of the model s implementation, computational resources, and configurations of hyper-parameters over all 13 datasets, we refer readers to Appendix I.2, I.3. ... We run L-STEP 5 times with different random seeds from 0 to 4 and report the average metric score. |